Over the past few years, my workflow as an engineer has changed a lot.
I went from the occasional autocomplete to using AI for ideation, debugging...
For further actions, you may consider blocking this person and/or reporting abuse
individual gains are obvious but team costs are subtle - that's what makes them dangerous to catch. in my experience the tell is whether teammates can still explain their decisions to each other, or just say 'AI suggested it.'
team knowledge sharing remains core I agree
yeah and the gap between "i know how this agent works" and "anyone on my team could debug it" is wider than most people track until someone goes on leave
It seems to me that you've really stumbled upon something profound here. The more we rely on AI, the more we risk losing sight of what makes us a team in the first place - our shared humanity. I love that he's proposing these practical exercises, like building a shared mental model, to keep our feet on the ground. It's such a simple yet vital reminder that AI is just a tool, not a replacement for the messy beauty of collaboration. Great Post Julien!
Thanks Aryan!
AI scales output. Teams still need to scale understanding.
100%
One very practical solution (that I also work on now) is to stop treating AI conversations as disposable personal artifacts of the team members.
Many useful things happen inside AI-assisted workflows. Let's say, developer investigates a rendering bug. AI suggests several possible causes - the engineer rejects some, verifies one, a workaround or real fix is found.
The entire logical chain disappears inside a private chat window.
That is lost team memory.
A possible solution is to collect selected AI-assisted work into a shared searchable store: PostgreSQL + pgvector, or a dedicated vector database.
The goal is not to store every token forever, but to preserve useful engineering context - for that the context should be analyzed and compress into a useful informational artifact:
Including links to the ticket, PR, annotated commits history, logs
Then the team can ask questions like:
This effectively turns AI output from private productivity pieces into a collective team memory.
I like that idea and framing. Thanks for sharing this.
the 'AI usage in PR template' practice is one of those things that sounds like overhead until you realize your team has been debug fixing AI code for months without anyone knowing where the shortcuts were taken.
we added a similar note to our internal reviews. the pattern that emerged: AI was being used most in the parts of the codebase nobody wanted to touch. the less understood the module, the more AI assistance, the more drift. circular.
the 'debugging without AI, narrated aloud' session is the one i'd actually steal. curious: do you record those, or is the value mostly in the live observation?
You can host a live session with your team and use it as a knowledge sharing session while collectively debugging. If you can record it that's even better, and get a transcript so you can use it as context for future tooling if beneficial.
This is really good. Just stopping to ask some of those critical questions makes you think more like an engineer.
Exactly. Are there particular critical questions that have helped you as an engineer?
Let's just say I've learned to question outputs in detail. :D I like your ideas though and will definitely give them some more consideration.
Having the PR template include a section on how AI was used is great. I'm gonna implement this with my team!
Happy this post inspired you Rasmus. Sounds great!
It's a little added friction but that can compound into very useful insights over time.